@doomdragon6 it's just an embedding that contains information on what is considered to be tokens that should be negated (i.e. tokens that when negated improve the quality of an image). An embedding is created via textual inversion and in simple terms, just maps to existing tokens in a model. You reference it by name in the negative prompt, adjust the strength to your liking (i.e. emphasis). To install it, all you do is place it into the embedding folder under the webui folder, perhaps needing to relaunch the UI to get it to load for use. Using this, you may not need to add as much to your negative prompt, assuming the embedding handles negating most of the typical things you would negative yourself.

@Havoc Gotcha! So it's a way to enter on negative prompt instead of 70, haha. Thank you for the informative comment!

Hello @NinjaDanny ,

The point of negative embeddings is they are trained to encompass tokens and weights that are associated with poor quality images, thus when you use them in the negative prompt, they will push the resulting image away from poor quality results. What it learns is low quality may not easily or intuitively prompted by someone without greater knowledge of the model. You can use such embeddings on their own, or in conjunction with your own negative prompts without issue.

@Havoc How do I use it, do I just put it in stable-diffusion-webui\embeddings folder?

so what ? How do I use it

@Havoc Best answer ever! Thanks for explaining this. I was having a hard time understanding this till now.

You're supposed to put it in the negative prompts

Hello @tamaokaima , see https://github.com/civitai/civitai/wiki/How-to-use-models#textual-inversions

Did you figure it out? I cant use it either

There is a description.. Use it in your negative prompt

@aimanartdude  yes

im having this problem too, did you fix it?

Hello @void2258 , it should work in newer versions of Automatic1111, I cannot speak to support in other UIs or applications which may accept embeddings in general. There is a pt version provided if needed.

Did you find a fix for this? Im having the same issue

Put the file in the embeddings folder

I have difficulties in installing these models and do inference on my computer, does anyone has any advices?

@maxwellcarey7777 What form of SD are you running? I recommend Automatic1111. There are some great tutorials on installing. Automatic1111 has context menus for including the models in diffusions after you download them into the proper directory.

Not sure how much info you are after, but very simply, if you want to stop something specific, save a bunch of generated images with that thing you don't want, and train a textual inversion using those images, then when you use that textual inversion in the negative prompt, it should, in theory, reduce the likelihood of that coming up. I use Kohya-SS GUI personally, mostly for LoRA, but you can also use the one built into Automatic1111, Aitrepreneur has really good guides on youtube

@BillySnooze Sounds good. If that is all it takes, sounds good to me. I trust that youtuber very much. I wouldn't have ever been able to get SD downloaded if it wasn't for him.

Well, I'm confused as to what I'm supposed to type in for the descriptions of the images to describe the parts that are a problem, and I don't understand how stuff like EasyNegative has a trigger word to use that encapsulates every single bad image you don't want, and I am trying to learn more about this because I don't understand the difference between us describing the images in a detailed way the way we do with prompts, and how they get one singular trigger word for that.

@SkyLordAer when you train a textual inversion you're basically assigning a bunch of features, styles, etc. To a single term in the case of the negative prompt TIs, the name of the inversion is the trigger word

It makes sense to use multiple negative embeddings together since they do not all direct the model away from making the same kinds of mistakes, and I have not noticed any problems from using them together. I routinely use both EasyNegative and boring_e621 together.

post pics please

it does make hands look realy neat but it doesn't fully stick to my prompt, do you know how to fix that?

Might not work for you but check your WebUI command window for any lines regarding the embedding. I noticed yesterday what one embedding was blocked cause it could be "malicious". It won't show that in the UI itself.

Should just need to put EasyNegative in the negative prompt. Worked for me, but didn't like the way it changed my images so I ultimately removed it. I downloaded the safetensor file, not the pickletensor.

easy diffusion cant use this, i just uninstalled them and installing automatic 1111

Now Easy diffusion can use this

@formertwitteremp5140 how?

@astolforeuploaded look in the instructions. Also there is a folder called hypernetwork

use badhandv4

Go to your InvokeAI install location and put the easynegative.safetensors file in the embeddings folder. Restart Invoke if it was already running, then just put <easynegative> in your negative prompt.

Should only need one, I recommend the safetensor file, as that is, well, safer. If you grab both, they probably conflict with each other. Make sure you write EasyNegative in the negative prompt to activate it. I recommend also checking with and without the prompt. Personally I didn't like how it changed my images so I removed it.

@natewilliams thank you

I have same problem after I reinstalled Stable Diffusion yesterday.. It seems not to be compatible with the new Version of Stable Diffusion. In Fact, no Textual Inversion I have on my PC is working anymore.

.pt files can have python code. Safetensors does not. You SHOULD use safetensors if possible because it doesn't have python code in it meaning less chance of malicious exploits. They generate the same output.

.pt files should pretty much never be used nowadays. Anybody making resources should be aware of this. If they're not using safetensors it's safe to assume that A. they're dumb or B. they're malicious. Either way doesn't bode well for your experience.

@Spikeles Oh I see, thanks. Should I be worried that almost all of my texture inversions are .pt files? Is there any way to check them for malware?

@oceanic  Aw geez, thanks for letting me know.

@BladdyVaddy 
https://github.com/lopho/pickle_inspector
https://github.com/zxix/stable-diffusion-pickle-scanner

Check the "tabs", one is safetensors and one is in pt

Still need help?

@taiconan  yes)

@pavloafasfafaf click the little i next to the word textual inversion and it will take you to a page and show you how to use it.

I'd be willing to try and offer suggestions. (I presume you've already tried clipskip 1-2 and sliding Step and CFG values one at a time?) Are you including anything else in your negative prompt to warrant trying to use it at half strength? Some embeddings will simply shred each other(by accident usually, though I did have one weird instance where two normally compatible embeddings were oddly incompatible but only in a specific checkpoint) and can occasionally interact with certain LoRAs/Lycoris in unexpected ways because of what's happening under the hood(though usually rarer unless it targets the exact nature of the embedding).

EDIT: Ah, you may also want to make sure you've got the PT file in your embeddings folder and not the Safetensor(Click the Tab Just below the title for the alternate version). If you're using the safe-tensor version, I believe it's not supposed to be placed in your regular prompt like it's a Lora? If this is the case and you've not put it in your negative prompt, you're potentially DELIBERATELY telling it to generate a bad image as the Lora that is supposed to tell it "Make bad image" has been turned into a tool for Creation of darkness and evil instead of suppressing it~!🤣
Any additional details?

@taiconan ideally it should work with the safetensor file too tho. because unlike pt or ckpt files, which apparently have executable code, safetensors files dont and are a lot safer that way

@taiconan can confirm. if you have the safetensor in the embeddings folder, but do not actually put it into the negative prompt, you will generate the weirdest voids of images imaginable.

@Eleiyas Hey, I'm currently trying to use EasyNegative in the negative prompts field and it's always generating black voids for me, using a basic positive prompt like "1girl, girl in field". I have the .safetensors file in the embeddings folder as instructed. Should i try the .pt one instead?

heyo dude, if you never came right or found out, any .Pt or safetensors can be used in the embeddings folder. but for me some safetytensors just never actually worked on my SD-Web-UI so that could be it too

E̶a̶s̶y̶ ̶D̶i̶f̶f̶u̶s̶i̶o̶n̶ ̶d̶o̶e̶s̶n̶'̶t̶ ̶s̶u̶p̶p̶o̶r̶t̶ ̶e̶m̶b̶e̶d̶d̶i̶n̶g̶s̶ ̶a̶s̶ ̶f̶a̶r̶ ̶a̶s̶ ̶I̶ ̶k̶n̶o̶w̶

Edit: Nevermind, the Beta channel seems to support embeddings. Go in the web ui on Settings -> toggle on the switch "Beta channel". Put your embedding in the stable-diffusion-ui\models\embeddings folder (create it if it doesn't exist already) and restart Easy Diffusion

mt, so, how did u fix it?

are you using the .pt or .safetensors? the wrong file will hurt your outputs

You have to put it in NEG Prompt

@kintilian45 where is the .pt file?

ComfyUI users must place it at: C:\ComfyUI_windows_portable\ComfyUI\models\embeddings

@KotatsuAi here how to call this into promtp https://www.youtube.com/watch?v=ElmyIsvMblE

Yes, he does. On the top of the page, there are switches between versions. pick the EasyNegative_pt

Look under the title!

You need to add the file to the models/embeddings folder, then add the following text to your prompt.

embedding:easynegative

Textual inversions are custom made CLIP embeddings that embody certain concepts. Textual inversions can be referenced inside a prompt by use the following syntax: embedding:name where name is the name of the embedding file.

Text Prompts - ComfyUI Community Manual (blenderneko.github.io)

no

you can use 'bad X' with SDXL I made it with the images of the correct size for normalized output

curious myself about this as well. I would like to use this, but don't want to overcomplicate my neg prompts. Any way to determine what's 'inside'? Thanks!

i have the same problem with this and other textual inversion files, i see a report of this problem in the release post of automatic1111 1.8, still no replies: 1.8.0 · AUTOMATIC1111/stable-diffusion-webui · Discussion #15073 (github.com)

edit: some days after I installed comfyui, and works better in my 4GB VRAM pc

@discco90279 So than it's either run a previous version of SD or wait for them to fix the issues. Although i see on a lot of forums saying many people use older or even the original version of Automatic1111 because of these updates

@jamesliebig106 I might have found a solution. I found this on youtube, Mastering Stable Diffusion: Common Errors and Easy Fixes (youtube.com),let me know if it works for you.

@junfon1234 One of the solutions that he demonstrated did fix the issue by switching the use to dedicated ram instead of vram, although thank goodness i have good amount of ram since it uses a sketchy amount of it haha

@jamesliebig106 Happy to hear I am not the only one having the problem. Not really happy though.

No embeddings works for me anymore. Not sure when it started, but I noticed it after installing control net comfyui. Never got comfyui to work, so it is not on the new install.

Now running on a clean new install of SD. But still, each time I add an embedding, it stops with a

*** Arguments: ('task(kqxk7b6k03cfeup)', <gradio.routes.Request object at 0x0000019FFB2810C0>, ' dakotaskyev2', '', [], 20, 'DPM++ 2M Karras', 1, 1, 7, 512, 512, False, 0.7, 2, 'Latent', 0, 0, 0, 'Use same checkpoint', 'Use same sampler', '', '', [], 0, False, '', 0.8, -1, False, -1, 0, 0, 0, False, False, 'positive', 'comma', 0, False, False, 'start', '', 1, '', [], 0, '', [], 0, '', [], True, False, False, False, False, False, False, 0, False) {}

Traceback (most recent call last):

The youtube video, seem to fix a start problem, but my SD runs fine, until I add aen embedding to the prompt.

Edit:

I downloaded version 1.7.0 zip from https://github.com/AUTOMATIC1111/stable-diffusion-webui/releases/tag/v1.7.0

Unpacked it in a new folder, symlinked my models folder and copied my embeddings to the new install folder.

And it is running and do not mind me adding embeddings to prompts.

Nah, dude. You have to put it into negative.

@pedrosapiens then why do most of the below examples don't have it in their prompts?

@JohnRohan Ya, it would be nice if the example prompts weren't inaccurate so much of the time. I AGREE.

I have the same question. My guess is, some people just simply copy and paste and negative prompts, without learning about textual inversion? Maybe I am wrong. But I am so curious about this

I think it is some kind of bug... unless Confy allows it... I am running Forge and sure can't use it...

You can use this online conversion tool from huggingface to convert any SD 1.5 embedding to an SDXL version: Embedding Converter - a Hugging Face Space by FoodDesert

@BladeRunner1988 Huh! What a handy tool! Thank you very much!

@BladeRunner1988 amazing and extremely clutch comment. W. Sigma Chad.

@BladeRunner1988 you should get sucked good an hard thru your jorts

@BladeRunner1988 you are a legend. sending buzz

anyone can share the converted embedding? Embedding converter on huggingface doesnt work anymore.

they're not - it's sd1.5 - on the backend you see an error saying it's being skipped

@myk3 They're not? Explain this then. https://civitai.com/images/18095423 also the 'extremely detailed' Lora that's advertised as SDXL, and apparently also used to produce the image above, doesn't work with pony, and gives an error stating that it's SD1. Does this site use metadata to list the assets used in an uploaded image's creation, or do people just manually enter the assets? The latter would explain why stuff that was apparently used to make an image, doesn't/can't. Otherwise, there's some lesser used program, or the civitai site image generator allows mixing of loras and models.

They might be using multi-model workflows.

if you already have krita set up with the ai features you just download it and place it in the embedding map

write as negative prompt: (easynegative: "choose a value")

Enter values ​​like those in the reference image

go to the textual inversion tab in your SD ui and reload the page, it should show up. then add it into your negative prompt and put <> around it

someone put a link in civit for converting the .pt (ds1.5) to .safetensore (sd1) I will put it here, I haven't used it yet

https://huggingface.co/spaces/FoodDesert/Embedding_Converter

same here

@JinZhao173 same again

Adds detail, removes noise and artifacts and improve lighting a bit. Its a strange "all in one" thing, that works with most models, even to a PONY level (at least for the pony Diffusion v6XL one). Illustrious is a point where I dont think it works anymore.

Hello! It seems embedding converter doesn't work, could you please share file which will work with ponydiffusion? Or is there a different way to convert embedding.

@Metakirin i wanted to ask for this too

Hi, sorry I am new to AI generate, but isn't the embed need to be used with the trigger word in the negative prompts? Your example's negative prompts doesn't have them, is it automatic applied as long as you put it in resource?

You can just type the name of the embedding in the negative window and it works. Like LORAs just without the "<name:1>" and only "name" or "(name:0.9)" if you want.

After doing test locally, I disagree. Its strange, but it still somehow works. Its not a lora and mainly used for adding more "juice" in an easier way, so I dont think its relevant what type of model you use (it works even with most Pony models, just not Illustrious from my testings).

When I stopped using it, the pictures were different. I also first thought it was just a relic of the past, but it still works.

@anotherstabledif1714 When you removed it, did you remove a trigger word or change any part of the prompt? That could have changed the output

colocalo en la carpeta de textual inversion y ya deberia aparecerte

@tanuki1994 eso hice pero a la hora de usarlo con un modelo poni no me aparece

Qué modelo estás usando? si es diferente a SD1.5 quizás no te aparezca porque el archivo no está guardado en formato .safetensor (tardé más de lo que me gustaría admitir en darme cuenta lol). Algunas respuestas a comentarios de este embedding tienen el enlace a la página, pero parece que no funciona (al menos a mi T.T)

@SlyCapricorn yo utilizo este https://civitai.com/models/605045/anitoon-pony-xl

I think people only use it because civitai does not hide it when using pony models. Negative hands, the other popular SD 1.5 embedding, gave me detached hands under pony and I had to drop it.

Somebody made a backup! https://huggingface.co/spaces/Quiho/Embedding_Converter-BackUp

The converter works now and it fixed the compatibility problem, so thanks!

Great guide. That said, I'm using CyberIllustriousXL (a merge of Cyberrealistic XL and Illustrious) and the unmodified version of EasyNegative still does a great job.

Thanks, I was having the same issue

where do you find the advanced generation mode?

@s2kim2022 It's the little gear all the way on the right when adding a new resource, to the right of relevance and filters.

EasyNegative was specifically trained to reduce overly idealized, anime-style beauty, especially big eyes, smooth skin, cute expressions, curvy body proportions, etc. These features are common in female characters, especially in anime or stylized models. So when it is applied, it can sometimes overcorrect and suppress masculine features too, or confuse the model's balance resulting in male characters being feminized.

As for why adding female into the negative prompt does not work, it is mainly because EasyNegative is not the same as just adding negative prompts. It is an embedding, meaning it subtly alters the model's understanding of the prompt at a deeper level. Saying "woman "in the negative prompt does not cancel out EasyNegative’s influence, because they are working on different layers of prompt interpretation.

I hope this answers your question!

@Hyokkuda does that mean I shouldn't use easynegative for anime generations

@walnutseal1461, I already shared my thoughts on whether or not you should use it in a comment here about three months ago. I suggest you try it both with and without. Sometimes, it works best when paired with ADetailer rather than globally.

我也有这个问题，请问你现在解决了吗？

@2364777293978 没有哦 还在摸索

@w858616我解决了，放在ComfyUI/models/embeddings就行了

@2364777293978 我放入这个文件夹之后该如何使用它，直接在提示词输入框打easynagetive吗？

@1241521519244 对的 没报错就是成功了

@2364777293978 好的 谢谢